JP5737725B2 - Localization and highly limited image reconstruction methods - Google Patents

Abstract

An image reconstruction method includes reconstructing a composite image of a subject using a conventional reconstruction method. The composite image employs the best information available regarding the subject of the scan and this information is used to constrain the reconstruction of a highly undersampled image frames or improve the SNR of image frames. A blurred and normalized weighting image is produced from image frame data, and this normalized weighting image is multiplied by the composite image.

Description

(Cross-reference of related applications)
This application is filed on Feb. 19, 2007, and is entitled US Patent Provisional Application No. 60/60, entitled “Localized and Highly Constrained Image Reconstruction Method”. Based on 901,728.

(Federal government commissioned study)
This invention was made with government support under grant number HL72260 approved by the National Institutes of Health. The US government has certain rights in the invention.

(Background of the Invention)
The field of the invention is medical imaging methods, in particular methods for reconstructing images from acquired image data.

Magnetic resonance imaging (MRI) creates images using the nuclear magnetic resonance (NMR) phenomenon. When a substance such as human tissue is exposed to a uniform magnetic field (polarizing magnetic field B ₀ ), the individual magnetic moments of the spins within the tissue attempt to align with this polarizing magnetic field, but around it their inherent Larmor Precess in random order with frequency. When this material or tissue is exposed to a magnetic field (excitation magnetic field B ₁ ) in the xy plane and close to the Larmor frequency, the net alignment moment M _z rotates with respect to the xy plane, Or “tilt” to create a net transverse magnetic moment M _t . After the signal is output by the excited spin and the excitation signal B ₁ is finished, this signal can be received and processed to form an image.

When creating an image using these signals, magnetic field gradients (G _x , G _y , and G _z ) are used. Typically, the area to be imaged is scanned by successive measurement cycles in which these gradients vary according to the particular positioning method used. In the industry, each measurement is called a “view”, and the number of views determines the quality of the image. The resulting received NMR signal or set of views or k-space sample numbers is digitized and processed to reconstruct the image using one of many well-known reconstruction techniques. The total scan time is determined in part by the length of each measurement cycle or “pulse sequence” and in part by the number of measurement cycles or views acquired for an image. There are many clinical applications where the total scan time for a given resolution and SNR image is important, and as a result, many improvements have been made to take this objective into account.

  Work has been done using projection reconstruction methods to acquire MRA data, and this method has been used again as disclosed in US Pat. No. 6,487,435. The projection reconstruction method does not sample k-space with a linear (Cartesian) scan pattern as shown in FIG. 2A as is done in the Fourier imaging method, but as shown in FIG. Sample k-space with a series of views that sample radial lines extending outwardly from the center of the space. The number of views required to sample the k-space determines the length of the scan, and if an insufficient number of views are acquired, streak artifacts are present in the reconstructed image. Created. The technique disclosed in US Pat. No. 6,487,435 obtains continuously undersampled images in an interleaved view and shares surrounding k-space data between successive images. By doing so, such streaking is reduced.

  In a computed tomography (“CT”) system, an X-ray source projects a fan beam and is collimated (collimated) so that the beam lies within the XY plane of a Cartesian coordinate system called the “imaging plane”. ) The X-ray beam passes through an imaging target such as a medical patient and strikes the radiation detector array. The intensity of the transmitted radiation depends on the attenuation of the X-ray beam by the object, and each detector produces a separate electrical signal that is a measurement of the beam attenuation. The attenuation measurements from all detectors are obtained individually to create a so-called “transmission profile”.

  The x-ray source and detector array in a conventional CT system rotates over the gantry in and around the imaging plane so that the angle at which the x-ray beam intersects the object always changes. The transmission profile from the detector array at a given angle is called the “view” and “scan” of the object, with different angular orientations during one revolution of the x-ray source and detector. Comprises a set of views made in In a two-dimensional scan, the data is processed to construct an image that corresponds to the two-dimensional slice obtained through the object.

  Similar to MRI, there are many clinical applications for X-ray CT where scan time is important. For example, in time-resolved angiography, a series of image frames (image frames) are acquired when the contrast agent flows into the region of interest. Each image is acquired as fast as possible to obtain a snapshot depicting the influx of contrast agent. This clinical application is particularly challenged when imaging coronary arteries or other blood vessels that require heart rate gating to suppress motion artifacts.

  For example, there are two methods used to reconstruct an image from a set of acquired k-space projection views, as described in US Pat. No. 6,710,686. In MRI, the most common method is to regrid k-space samples from a position on a trajectory sampled in the radial direction to a Cartesian grid. The image is then reconstructed by a 2D or 3D Fourier transform of the rescaled k-space sample. A second way to reconstruct the MR image is to transform the radial k-space projection view to Radon space by first Fourier transforming each projection view. Images are reconstructed from the projected views by filtering and backprojecting these signal projected views to the field of view (FOV). As is well known in the art, if the acquired signal projections are insufficient to satisfy the Nyquist sampling theorem, streak artifacts will occur in the reconstructed image.

  A widely practiced method for reconstructing images from two-dimensional X-ray CT data is referred to in the art as the filtered back projection technique. This backprojection process is substantially similar to that used in the MR image reconstruction described above, according to which the attenuated signal measurements acquired during the scans represent the brightness of the corresponding pixels in the display. It is converted into an integer called “CT value” or “Hansfield value” used for control.

（ここで、Ｓ_nは、Ｎ個の画素を有する投影経路内のｎ番目の画素に分配される信号値である。） A standard backprojection method used for both MRI and X-ray CT is shown in FIG. Each acquired signal projection profile 10 is backprojected onto the FOV 12 by projecting each signal sample 14 along the projection path indicated by the arrow 16 through the FOV 12 within the profile 10. In projecting each signal sample 14 to the FOV 12, no a priori information about the object to be imaged is used, the signal in the FOV 12 is homogeneous, and the signal sample 14 is passed through each pixel through the projection path. Are evenly distributed. For example, in FIG. 3, the projection path 8 for a single signal sample 14 in a single projection profile 10 is shown as the projection path 8 passes through N pixels in the FOV 12. The signal value (P) of the signal sample 14 is evenly divided among these N pixels.

(Here, _Sn is a signal value distributed to the nth pixel in the projection path having N pixels.)

  Clearly, the assumption that the FOV 12 backprojection signal is equal is incorrect. However, as is well known in the art, this false assumption can be made if a certain correction is made to each signal profile 10 and a sufficient number of profiles are acquired with a corresponding number of projection angles. The resulting error is minimized and image artifacts are suppressed. A typical filtered backprojection method for image reconstruction requires 400 projections for a 256 × 256 pixel 2D image and 203,000 for a 256 × 256 × 256 pixel 3D image. Individual projections are required.

  Recently known in the art as “HYPR” and co-pending US patent application Ser. No. 11 / 482,372 (filed Jul. 7, 2006, entitled “Highly Limited Image Reconstruction”). A new image reconstruction method described in “Highly Constrained Image Reconstruction Method” has been disclosed, the contents of which are incorporated herein by reference.A priori of objects to be imaged by the HYPR method The composite image is reconstructed from the acquired data to provide the information that is then used to highly limit the image reconstruction process.HYPR is a magnetic resonance imaging (MRI) method. ), X-ray computed tomography (CT), positron emission tomography (PET), single photon emission computed tomography (SPECT) and digital tomosynthesis It can be used in a number of different diagnostic imaging methods including DTS.

  As shown in FIG. 1, for example, when a series of time-resolved images 2 are acquired in a dynamic study, each image frame 2 can be reconstructed using a very limited set of acquired views. However, each such view set is interleaved with the views acquired for other image frames 2, and after a large number of image frames are acquired, a sufficient number of different views are used to generate the HYPR method. The high-quality composite image 3 for use according to the above can be reconstructed. The composite image 3 generated by using all of the interleaved projections is fairly high image quality, and this high image quality is transferred to the image frame by using the highly limited image reconstruction method 4 of the present invention. The Image frame 2 can also be acquired in a dynamic test where the dose (eg, X-rays) or exposure time (eg, PET or SPECT) is reduced for each image frame. In this case, the composite image is generated by accumulating or averaging measurements from a series of acquired image frames. The highly constrained reconstruction 4 of each image frame 2 conveys the high SNR of this composite image to the resulting reconstructed image.

（ここで、Ｓ_nは、再構成されている画像フレーム内の画素ｎにおける逆投影された信号の大きさであり、Ｐは、逆投影されている投影プロファイルにおける信号サンプル値であり、Ｃ_nは、逆投影経路に沿ったｎ番目の画素における先験的な合成画像の信号値である。）
合成画像は、走査中に取得されたデータから再構成され、フィールド・オブ・ビューで構造体を表す、他の取得画像データだけでなく画像フレームを再構成するために用いられる合成画像を含んでもよい。数式（２）の分子は、合成画像において、対応する信号値を用いた各画素に重み付けし、分母は、全逆投影信号サンプルが、画像フレームに対する投影総和を反映し、合成画像の総和により乗算されないように、その値を正規化している。 The discovery of the HYPR method is that if a priori information of the signal contour in the FOV 12 is used in the reconstruction process, a good quality image can be generated using significantly fewer projection signal profiles. For example, referring to FIG. 4, it can be seen that the signal contour of FOV 12 includes structures such as blood vessels 18 and 20. In practice, when the backprojection path 8 penetrates these structures, a more accurate distribution of the signal sample 14 to each pixel is achieved by weighting the distribution as a function of the known signal contour at that pixel location. Is done. As a result, the majority of signal samples 14 are distributed in the backprojected pixels that intersect the structures 18 and 20 in the example of FIG. For a backprojection path 8 with N pixels, this highly restrictive backprojection can be expressed as:

(Where S _n is the magnitude of the backprojected signal at pixel n in the reconstructed image frame, P is the signal sample value in the backprojected projection profile, and C _n Is the a priori composite image signal value at the nth pixel along the backprojection path.)
The composite image may be reconstructed from the data acquired during the scan, including the composite image used to reconstruct the image frame as well as other acquired image data representing the structure in the field of view. Good. The numerator of equation (2) weights each pixel using the corresponding signal value in the composite image, and the denominator is the total backprojection signal sample reflecting the projection sum for the image frame, multiplied by the sum of the composite image The value is normalized so that it is not.

It should be noted that this normalization can be done separately for each pixel after backprojection, but for many clinical uses it is much easier to normalize projection P before backprojection. is there. In this case, the projection P is normalized by dividing by the corresponding value P _C in a projection through the composite image at the same view angle. The normalized projection P / P _C is then backprojected and the resulting image is then multiplied by the composite image.

（ここで、総和（Σ）は、再構成される画像フレーム内の全投影であり、特定のラドン平面内のｘ、ｙ、ｚ値は、その平面に対する適正なｒ、θ、φ値における投影プロファイル値Ｐ（ｒ，θ，φ）を用いて算出される。Ｐ_c（ｒ，θ，φ）は、合成画像からの対応投影プロファイル値であり、Ｃ（ｘ，ｙ，ｚ）_r,θ_,φは、（ｒ，θ，φ）での合成画像値である。） A highly limited backprojection 3D implementation is shown in FIG. 5 for a single 3D projection view characterized by view angles θ and φ. This projection view is backprojected along the axis 16 and extends into the Radon plane 21 at a distance r along the backprojection axis 16. Instead of a filtered backprojection in which the projection signal values are filtered and distributed evenly along the axis 16 into a continuous Radon plane, the projection signal values are used in the Radon plane using information in the composite image. 21. The composite image in the example of FIG. 5 includes blood vessels 18 and 20. Weighted signal contour values are placed at image positions x, y, z in the Radon plane 21 based on the intensity at corresponding positions x, y, z in the composite image. This is a simple multiplication of the voxel value of the corresponding composite image and the back-projected signal profile value P. This product is then normalized by dividing this product by the projection profile value from the corresponding image space projection profile formed from the composite image. The equation for 3D reconstruction is:

(Where the sum (Σ) is the total projection in the reconstructed image frame, and the x, y, z values in a particular Radon plane are the projections at the appropriate r, θ, φ values for that plane. Calculated using the profile value P (r, θ, φ), where P _c (r, θ, φ) is a corresponding projection profile value from the composite image, and C (x, y, z) _r, θ. _, φ are composite image values at (r, θ, φ).)

  The HYPR method works well under certain circumstances. First, HYPR works best with “sparse data sets, where the object of interest in the field of view occupies a small portion of the image space. A range of images with very sparse or perfect spatio-temporal correlation. Within this, the HYPR reconstruction method provides a substantially accurate reconstruction: when the sparse spatio-temporal correlation is reduced, the accuracy of the HYPR reconstruction is reduced due to the influence of signals elsewhere in the field of view with different temporal behavior. The local signal can change.

(Summary of the Invention)
The present invention is a new method for generating medical images, in particular an improved method for generating images using the HYPR reconstruction method. The initial frame obtained from the limited data is used to generate a normalized weighted image, which is multiplied with a high quality composite image to generate a high quality image frame. The normalized weighted image is generated by smoothing the initial image frame with a filter and dividing the smoothed image frame by a similar smoothed version of the composite image. If the initial image frame is limited by undersampling, the version of the composite image used to generate the normalized weighted image is “limited” to the same undersampled view embedded in the initial image frame. .

  The general object of the present invention is to improve the quality of image frames generated using limited data. The data is limited in the sense that an image frame with a low signal-to-noise ratio (“SNR”) is generated and / or generated from a view where the data undersamples the acquisition space (eg, k-space or radon space). It is limited in the meaning. In any case, the high quality of the composite image is incorporated into the image frame according to the present invention.

各ピクセルの強度Ｉ（ｘ，ｙ）は、均等に限定されたカーネルを有するＩ（ｘ，ｙ）のコンボリューションであるピクセルｘ，ｙの最近傍の総和になる。これは再構成された画像フレームのローパスフィルタ・バージョンと等価である。さらに詳細には、画像フレームのローパスフィルタ・バージョンは、合成画像のローパスフィルタ・バージョンで割算されることにより正規化され、次に、結果として得られた正規化重み付け画像は最高解像度の合成画像と乗算される。ローパスフィルタリング演算はラドン空間またはｋ空間内で実行でき、または画像空間内で実行できる。この局所化ＨＹＰＲ再構成は、局所ボリューム、またはセグメント、周囲の各ピクセル内の各信号を使用し、視野内の他の場所で生成された信号のクロストークが除去される。 Another general object of the present invention is to improve the performance of the HYPR reconstruction method when non-sparse data sets are acquired. In order to reduce the effect of distance signals in the field of view (FOV) reconstructed using the HYPR method, the field of view is divided into smaller segments that are individually reconstructed. Only those portions of the projection view that project forward through the small segment to be reconstructed are used for HYPR reconstruction of that segment. The reconstructed small image segments are then combined to provide a reconstructed image of the entire FOV. This is illustrated in FIG. 19, where the projections P1, P2,. . . Pn is backprojected onto a limited segment S (x, y) of the entire FOV 5. As the size of each segment S (x, y) decreases and the number of segments increases, the “localized” HYPR reconstruction looks as a division into small regions surrounding each image pixel I (x, y). be able to. A value for intensity I (x, y) is obtained by summing all of the backprojected values P1 _lim to Pn _lim in segment S (x, y).

The intensity I (x, y) of each pixel is the sum of the nearest neighbors of pixel x, y, which is a convolution of I (x, y) with an evenly limited kernel. This is equivalent to a low-pass filter version of the reconstructed image frame. More specifically, the low-pass filter version of the image frame is normalized by dividing by the low-pass filter version of the composite image, and then the resulting normalized weighted image is the highest resolution composite image. And multiplied. The low pass filtering operation can be performed in Radon space or k space, or can be performed in image space. This localized HYPR reconstruction uses the local volume, or segment, each signal in each surrounding pixel, and the crosstalk of the signal generated elsewhere in the field of view is removed.

  Another object of the present invention is to avoid the need for Radon transforms or backprojection operations. Localized HYPR reconstruction can be performed using a series of low-pass filtering operations, division operations and multiplication operations. These operations can be performed in acquisition space (eg, Radon space or k-space) or image space.

  Another object of the present invention is to provide a HYPR reconstruction method that can be used in data acquired by many different imaging modalities and many different sampling patterns. This method can be used to reconstruct an image from data acquired by an MRI system, X-ray CT system, PET scanner or SPECT scanner. When used in an MRI system, k-space data can be acquired using any sampling trajectory. This method can also be used to enhance radiographic images.

  The above and other objects and advantages of the present invention will become apparent from the following description. In the description, reference is made to the accompanying drawings that form a part hereof, and in which is shown by way of illustration one preferred embodiment of the invention. Such embodiments, however, do not necessarily represent the full scope of the invention, and reference should be made therefore to the claims herein for interpreting the scope of the invention.

Detailed Description of Preferred Embodiments
The present invention can be applied to many different medical imaging modalities and many different clinical applications. In the preferred embodiment described below, a projected view of an object to be imaged is acquired using an MRI system, but the present invention may be used in other imaging modalities and sampling other than a radial projection view. A trajectory may be acquired.

  With particular reference to FIG. 6, the preferred embodiment of the present invention is employed in an MRI system. The MRI system includes a workstation 110 having a display 112 and a keyboard 114. The workstation 110 includes a processor 116 that is a commercially available programmable machine running a commercially available operating system. The workstation 110 provides an operator interface that allows scan instructions to be entered into the MRI system.

  The workstation 110 is coupled to four servers: a pulse sequence server 118, a data acquisition server 120, a data processing server 122, and a data storage server 23. In the preferred embodiment, data storage server 123 is implemented by workstation processor 116 and associated disk drive interface circuitry. The remaining three servers 118, 120 and 122 are executed by separate processors mounted in a single enclosure and interconnected using a 64-bit backplane bus. The pulse sequence server 118 employs a commercially available microprocessor and a commercially available 4 communication controller. Both the data acquisition server 120 and the data processing server 122 employ the same commercially available microprocessor, and the data processing server 122 further comprises one or more array processors based on a commercially available parallel vector processor.

  The workstation 110 and each processor of the servers 18, 20 and 22 are connected to a serial communication network. This serial network conveys data downloaded from the workstation 110 to the servers 118, 120 and 122, as well as tag data communicated between the servers and between the workstation and the server. In addition, a high speed data link is provided between the data processing server 122 and the workstation 110 to transmit image data to the data storage server 123.

The pulse sequence server 118 functions to operate the gradient system 124 and the RF system 126 in response to program elements downloaded from the workstation 110. The gradient waveform required to perform the specified scan is generated and provided to the gradient system 124, which excites the gradient coil in the assembly 128 to use the magnetic field gradient used for position encoding of the NMR signal. G _x , G _y , and G _z are generated. The gradient coil assembly 128 forms part of a magnet assembly 130 that includes a polarizing magnet 132 and a whole body RF coil 134.

  An RF excitation waveform is provided by the RF system 126 to the RF coil 134 to perform the specified magnetic resonance pulse sequence. The responsive NMR signal detected by the RF coil 134 is received by the RF system 126 and amplified, demodulated, filtered and digitized under the command of the command generated by the pulse sequence server 118. The RF system 126 includes an RF transmitter that generates a wide range of RF pulses used in MR pulse sequences. The RF transmitter generates RF pulses of the desired frequency, phase, and pulse amplitude waveform in response to scan instructions and commands from the pulse sequence server 118. The generated RF pulse can be applied to the whole body RF coil 134 and can be applied to one or more local coils or coil arrays.

また、受信したＮＭＲ信号の位相も求めることができる。

The RF system 126 also includes one or more RF receiver channels. Each RF receiver channel includes an RF amplifier that amplifies the NMR signal received by the connected coil, and a quadrature detector that detects and digitizes the I and Q quadrature components of the received NMR signal. The magnitude of the received NMR signal can thus be determined at any sampling point by the square root of the square sum of the I and Q components,

Further, the phase of the received NMR signal can also be obtained.

  The pulse sequence server 118 optionally receives patient data from the physiological acquisition controller 136. The controller 136 receives signals from a number of different sensors connected to the patient, such as ECG signals from the electrodes or respiratory signals from the bellows. The pulse sequence server 118 typically uses such signals to synchronize or “gate” the performance of the scan to the patient's breath or heartbeat.

  The pulse sequence server 118 also connects to a scan room interface circuit 138 that receives signals from various sensors related to the condition of the patient and the magnet system. It is also through the scan room interface circuit 138 that the patient alignment system 140 receives commands to move the patient to the desired position during the scan.

  It should be apparent that the pulse sequence server 118 provides real-time control of MRI system elements during the scan. As a result, it is necessary for the hardware element to operate with program instructions that are executed in a timely manner by the runtime program. The command component of the scan instruction is downloaded from the workstation 110 in the form of an object. The pulse sequence server 118 includes programs that receive these objects, and converts these objects into objects that are employed by the runtime program.

  The data acquisition server 120 receives digitized NMR signal samples generated by the RF system 126. Data acquisition server 120 operates in response to command components downloaded from workstation 110, receives real-time NMR data, and provides buffer storage so that data is not lost due to data overruns. Depending on the scan, the data acquisition server 120 simply passes the acquired NMR data to the data processor server 122. However, in scans that require information derived from acquired NMR data to control further performance of the scan, the data acquisition server 120 is programmed to generate such information and communicate it to the pulse sequence server 118. Is done. For example, during pre-scan, NMR data is acquired and used to calibrate the pulse sequence performed by the pulse sequence server 118. In addition, navigator signals can be acquired during a scan and used to adjust the operating parameters of the RF system or gradient system, or to control the view order in which k-space is sampled. The data acquisition server 120 can also be employed to process NMR signals that are used to detect the arrival of contrast agents during an MRA scan. In all of these examples, the data acquisition server 120 acquires NMR data and processes it in real time to generate information used to control the scan.

  The data processing server 122 receives the NMR data from the data acquisition server 120 and processes it according to the command component downloaded from the workstation 110. Such processing includes, for example, Fourier transforming unprocessed k-space NMR data to generate a two-dimensional or three-dimensional image, applying a filter to the reconstructed image, and acquired NMR data. Performing back-projection image reconstruction, calculating functional MR images, calculating motion or flow images, and the like. As described in further detail below, the present invention is executed by the MRI system in response to a program executed by the data processing server 122.

  The image reconstructed by the data processing server 122 is transmitted again to the workstation 110 and stored. Real-time images are stored in a database memory cache (not shown) from which images can be output to an operator display 112 or display 142 that is located near the magnet assembly 130 and used by the attending physician. Batch mode images or selected real-time images are stored in a host database on disk storage 144. When such an image is reconstructed and transferred to the storage, the data processing server 122 notifies the data storage server 123 on the workstation 110. An operator can use the workstation 110 to store images, generate film, or send images to other facilities over a network.

  In order to implement the preferred embodiment of the present invention, NMR data is acquired using a pulse reconstruction as shown in FIG. 7, which is projection reconstruction or radial. This is a fast gradient recalled echo pulse sequence in which a selectively, asymmetrically cut sinc rf excitation pulse 200 is generated in the presence of a slice selection gradient 202. This pulse sequence can be used to obtain a single 2D slice by sampling on a single k-space circular slice, or this pulse sequence can be shown as 204, 206 and 208 in FIG. As can be used, it can be used to sample multiple circular k-space planes. When multiple 2D slices are acquired, the gradient 202 is the selected gradient that follows the phase-encoded gradient lobe 210 and the reverse polarity rewinder gradient lobe 212. This axial phase encoding gradient 210 is stepped through the values during scanning to sample from each of the 2D k-space slices 204, 206 and 208.

  Two in-plane readout gradients 214 and 216 are performed during acquisition of the NMR echo signal 218 to sample k-space along a radial trajectory in the 2D plane 204, 206 or 208. These in-plane gradients 214 and 216 are perpendicular to the axial gradient and they are orthogonal to each other. As described in more detail below, during scanning, these in-plane gradients are stepped through a series of values to rotate the view angle of the radial sampling trajectory. Each in-plane read gradient precedes pre-fading gradient lobes 220 and 222, followed by unwind gradient lobes 224 and 226.

To implement another preferred embodiment of the present invention, a pulse sequence that acquires data is used as the 3D radial projection shown in FIG. This sequence is performed on the MRI system described above equipped with a fast gradient subsystem (40 mT / m maximum amplitude and 150 T / m / sec maximum slew rate). Full echo or partial echo readout can be performed during the data acquisition window 199. When partial echo is selected, the lower half of k-space (k _z <0) is only partially acquired. For large FOVs in all directions, non-selective radio frequency (RF) pulses 209 can be used to generate transverse magnetization throughout the image FOV.

A gradient recalled NMR echo signal 203 is generated by spins in the excited FOV and acquired in the presence of three readout gradients 201, 211 and 207. Since no slab selection gradient is required, the readout gradient waveforms G _x , G _y and G _z have a similar shape. This symmetry is only broken by the need to invalidate the sequence and is achieved by performing a detuning gradient lobe 205. G _x , G _y and G _z readout gradients 201 and 211 are rewound by respective gradient pulses 213 and 215 to achieve a stable state.

The readout gradient waveforms G _x , G _y and G _z are modulated during the scan to sample the radiation trajectory at various angles. The angular space is selected such that an even distribution of k-space sample points occurs at the peripheral boundary (k _max ) of the sampled k-space sphere. Several methods for calculating the distribution are known, but a method is used in which the projections are evenly distributed by sampling the sphere with a spiral trajectory at a constant path velocity and surface area range. This solution also has the advantage of generating a continuous sample path to reduce gradient switching and eddy currents. For a total of N projections, the equation for gradient amplitude as a function of the number of projections n is

  The readout gradient amplitude for the nth pulse sequence in this series of projections is given by equations (5), (6) and (7). It should be understood that n can be indexed in a monotonic order from 1 to N during the scan, although other orders are possible. As described below, the present invention allows the spherical k-space to be sampled with significantly fewer projection views, resulting in a shorter scan time.

  Those skilled in the art may use sampling trajectories other than the preferred linear trajectory that extends from a point on the perimeter boundary of k-space to the opposite point on the perimeter boundary of k-space through the center of k-space. Should be understood. As described above, one variation is to obtain a partial NMR echo signal 203 that samples along a trajectory that does not extend over the entire range of the sampled k-space volume. Another variation that is equivalent to a linear projection reconstruction pulse sequence is to sample along a curved path rather than a straight line. Such a pulse sequence is, for example, F.I. E. Boada et al., “Fast Three Dimensional Sodium Imaging” (MRM, 37: 706-715, 1997); V. Koladia et al., “Rapid 3D PC-MRA Using Spiral Projection Imaging (Helix Projection Imaging)” (Proc. Intl. Soc. Magn. Reson. Med. 13 (2005)) and J. G. It is described in “Spiral Projection Imaging: a new fast 3D trajectory” (Proc. Intl. Soc. Mag. Reson. Med. 13 (2005)) by Pipee and Koladia. The present invention is available in 3D versions as well as 2D versions of these sampling methods, and the use of the term “pixel” as used herein refers to a position in either a 2D or 3D image. It should also be understood that it is intended.

  The MRI system described above can be used in a wide range of clinical applications to acquire 2D or 3D projection view sets that are used to reconstruct one or more images. The image reconstruction method of the present invention is particularly useful in scans where one or more image frames are reconstructed using fewer projection views than the entire acquired projection view. In this application, image artifacts that normally occur due to undersampling are eliminated or suppressed.

  A first embodiment of an image reconstruction method relates to an MRI system for acquiring a two-dimensional projection view and reconstructing a series of image frames that depict an object over a period of time. In particular, referring to FIG. 9, a projection view set is obtained and a series of image frames is reconstructed therefrom, as indicated by process block 225. The number of projection views in each set is small (eg, 10 views) and is evenly distributed to sample k-space as uniformly as possible, as shown in FIG. Since the number of acquired projection views is small, this image frame can be acquired with a very short scan time, but since the k-space is highly undersampled, it was reconstructed using conventional methods. In any image, streak artifacts will occur.

  The next step, indicated by process block 227, is to combine the projected views obtained from the test subject and reconstruct the composite image. This generally includes a projected view that is acquired within a time window surrounding an acquired view of the current image frame and interleaved with the view of the current image frame. The composite image projection is a significantly larger number than the image frame dataset, which provides a more complete k-space sampling. As a result, the reconstructed composite image has a high signal-to-noise ratio (SNR) and striped artifacts are suppressed. In the preferred embodiment, this reconstruction includes re-sizing the combined k-space projection data to Cartesian coordinates and then performing an inverse two-dimensional Fourier transform (2DFT) to generate a composite image.

  As indicated by process block 229, the next step is to reconstruct the HYPR image frame according to the localization method of the present invention. There are a number of different ways to accomplish this, which are described in detail below with reference to FIGS.

  After the HYPR image frame is reconstructed, a test is made at decision block 243 to determine whether additional image frames should be generated. If so, the system loops back through process block 241 to reconstruct another composite image as indicated by process block 277. In a dynamic test of an object, for example, a series of image frames are acquired and reconstructed to display a certain temporal variation method of the object. Projection views acquired during this test are interleaved with each other so that these projection views become samples of different parts of k-space and the composite image is within a time window centered on the reconstructed HYPR image frame. Generated by combining projection views acquired in The width of this time window includes enough projected views to adequately sample k-space so that striped artifacts are suppressed and high SNR is maintained, but significantly reduces the temporal resolution of the image frame. Set to not include many projection views.

  As indicated by process block 245, when the last image frame has been reconstructed as determined at decision block 243, the reconstructed image frame is stored. Stored image frames can be displayed one at a time or can be displayed continuously to show how the subject changes during dynamic testing.

（ここで、ＳＮＲ_compositeは合成画像におけるＳＮＲであり、Ｎ_fは時系列の画像フレームの数であり、Ｎ_vは投影の際のオブジェクト画素数であり、Ｎ_pixは投影の際の画素数（例えば、２Ｄでは２５６個、または３Ｄでは２５６×２５６個）、Ｎ_pは画像フレームごとの投影数である。）。Ｎ_pが１０のオーダーである場合、再構成されたＨＹＰＲ画像フレームのＳＮＲは、ＳＮＲ_compositeによって決定される。 It can be demonstrated that the SNR of each reconstructed image frame is determined by the SNR of the composite image. The SNR is calculated as the ratio of the object signal level to the noise standard deviation in the object, and the CNR is calculated as the difference between the object signal level divided by the standard deviation of the background noise and the background signal level. SNR and CNR are generally limited by a combination of stochastic noise and noise due to streak artifacts. The stochastic component of SNR in the highly limited backprojection reconstruction of the present invention is given by:

(Where SNR _composite is the SNR in the _composite image, N _f is the number of time-series image frames, N _v is the number of object pixels during projection, and N _pix is the number of pixels during projection ( (For example, 256 in 2D, or 256 × 256 in 3D), and N _p is the number of projections per image frame.) If N _p is on the order of 10, the SNR of the reconstructed HYPR image frame is determined by the SNR _composite .

  The composite image may be acquired or reconstructed in a number of different ways depending on the clinical application. In the above embodiment of the present invention, the initial composite image is reconstructed from interleaved views acquired within a time window centered near the acquisition time of the reconstructed image frame. This is particularly applicable in the presence of substantial movement in the subject or substantial change in the subject during dynamic testing. As shown in process block 247 of FIG. 9, in other clinical applications, the initial composite image may also be acquired during a separate scan that is not time-constrained and the k-space is fully sampled. . Thereby, an anatomical image having a high resolution and a high SNR can be acquired and used as a composite image.

With particular reference to FIG. 10, the first step in reconstructing a HYPR image frame is to reproject the composite data set C _t from the composite image for the current image frame, as shown in process block 250. The current image frame consists of a small number of projection views acquired at the selected view angle, and the reprojected composite data set C _t is generated by reprojecting the composite image at these same view angles. .

In the next step, indicated by process block 252, the composite data set _Ct is smoothed. This is accomplished by Fourier transforming each projection view in the data set C _t and then multiplying the resulting k-space projection by the Fourier transform filter kernel described in detail below.

As indicated by process block 254, the composite image C _t is then reconstructed from the smoothed data set C _t . This is a conventional image reconstruction such as, for example, changing the grid of radiation k-space data to Cartesian coordinates and then performing an inverse Fourier transform. The resulting limited composite image C _t contains striped artifacts due to undersampling, but these striped artifacts are included in the next image frame because each contains the same view angle. It is substantially the same as the inside artifact.

As indicated by process block 256, the k-space projection view for the current image frame is also smoothed using a Fourier transform filter kernel. This is accomplished by multiplying each k-space projection view using the same filter described above for the composite data set C _t . The smoothed current image frame T is then reconstructed using the filtered k-space data set, as indicated by process block 258. As in the smoothed composite image _Ct , this is a standard image reconstruction and the fringes are generated by undersampling.

As indicated at process block 260, then, by dividing the current smoothed image frame T by the current smoothed composite image C _t, normalized weighting image T _w is generated. This is a direct division between pixels, except that the pixels in the filtered composite image C _t that are zero are initially set to a small value. By performing this division operation separately on the real and imaginary components of the complex pixel value, the phase information is maintained. As indicated at process block 262, then, by multiplying the complete composite image C and the normalized weighting image T _w, HYPR image frame T _H is generated. This is a direct multiplication between the corresponding pixels in the two images.

In this embodiment of the invention, the normalized weighted image is “smoothed” in k-space by a multiplication operation. In the second embodiment described next, this smoothing is performed in the image space by a filtering operation. In any case, the high SNR of the composite image, is modified by the normalized weighting image T _w, to display the object at the time between the current image scan frame data is obtained. The high SNR of the composite image is maintained along with time dependent information for the current image frame.

With particular reference to FIG. 11, in the second embodiment of the present invention, “smoothing” is performed in the image space, not the k-space described above. The first step in the second embodiment of the present invention is to regenerate the composite data set C _t for the current image frame, shown in process block 270. This is the same step as described above in process block 250. As indicated at process block 272, a “limited” composite image is then regenerated from the composite data set C _t using conventional image reconstruction methods.

As indicated in process block 274, the limited composite image C _t is then smoothed by filtering. More particularly, smoothing is a convolution process in the image space, and the limited composite image C _t is convolved with a filter kernel described in more detail below.

As indicated at process block 276, an image frame is then reconstructed from the limited projection views acquired for the current time during the scan. This is a conventional image reconstruction, and as a result, a striped pattern due to undersampling appears in the resulting image frame T and has a relatively low SNR. As indicated at process block 278, the image frame is smoothed by filtering. Filtering is performed in the image space for the limited composite image C _t using the same filter kernel by the same convolution process as described above.

As shown in process block 279, a normalized weighted image T _w is then generated by dividing the filtered image frame T by the filtered composite image C _t . This is a conventional division operation for dividing the corresponding pixel values in the filtered is filtered each pixel value in the image frame T and composite image C _t. Next, as shown in process block 280, by multiplying the complete composite image C in the normalized weighting image T _w, to produce the final HYPR image frame T _H.

  In the first two embodiments described above, a filter kernel or a Fourier transform of this kernel is used to smooth the image in image space or k-space. In a preferred embodiment, the 2D image is smoothed using a 7 × 7 square filter kernel in image space, and the k-space image data is smoothed using the Fourier transform sinc function of this kernel. When reconstructing a 3D image, a 9 × 9 × 9 cubic filter kernel or sphere kernel with equal weighting throughout the kernel is used in image space and smoothed in k-space using the Fourier transform of this kernel . The kernel size should be selected so that when smoothing is performed, the kernel does not contain a large amount of information from outside the object of interest (eg, a blood vessel). The kernel size should be as small as or slightly smaller than the dimensions of the object being tested, but its exact shape is not critical. Gaussian or other smooth kernels may also be used and the resulting function performed is low pass filtering. “You should add any or all of the information or teachings you got about the filter kernel that can be used in the present invention.”

In the first two implementations, the filtering is performed separately on the limited composite data or image C _t and the current image frame data or image. It is also possible to first perform a division operation in k-space or image space, and then perform a smoothing function on the result in either k-space (multiplication) or image space (convolution) It is. All striped patterns that occur after the division operation are smoothed when this embodiment is employed.

  In the above-described embodiments of the present invention, smoothing or filtering, operations are performed on k-space data or image space data. An equivalent but less desirable way to achieve the same result is to divide the image space field into smaller segments and reconstruct each small segment individually using the HYPR method. The segments are then combined to produce a complete reconstructed image. As described above, this method is equivalent to the smoothing method described above. This is because the results are theoretically the same when the segment size is reduced to 1 pixel. A preferred embodiment of this equivalent “segmented” HYPR method will now be described below.

  With particular reference to FIG. 13, any of the above pulse sequences may be used to acquire a series of image frames, as shown in process block 325. For example, these images may be acquired as part of a CEMRA dynamic test where a series of image frames are acquired from the vasculature when the contrast bolus enters the field of view. Each image frame is highly undersampled to increase the temporal resolution of the test, and the acquired projection view of each undersampled image frame is interleaved with the acquired projection view of the other image frame .

  As shown in process block 327, the next step is to reconstruct a high quality composite image for use in a highly limited reconstruction process. In this embodiment, the composite image combines all interleaved projection views from the acquired image frames into a single highly sampled k-space data set, which is conventionally used with the combined k-space data set. Is generated by performing image reconstruction. Conventional reconstruction may perform a filtered backprojection after the Fourier transform of each projection view, or perform a 3D Fourier transform after re-sizing k-space samples in the combined k-space data set to Cartesian coordinates. Good. In an alternative method, individual composite images may be acquired and reconstructed.

  Still referring to FIG. 13, a loop is then entered in which the image frames are reconstructed. Although a single image frame may be reconstructed, in general, the image frame reconstruction loop is repeated a number of times to reconstruct all image frames acquired during dynamic testing.

  As shown in process block 329, the first step in this loop is to reconstruct the image using the acquired highly undersampled image frames. Since this is a conventional image reconstruction and the acquired image frames are highly undersampled, the resulting image contains many striped patterns and other artifacts, thereby increasing its diagnostic value. There is a possibility to limit. The next step shown in process block 331 is to divide this conventional image frame into segments. As shown in FIG. 15, segmentation in the present invention means dividing a 3D FOV 300 into smaller 3D volumes, or segments 302.

  After the conventional image frame is divided into segments, a loop is entered in which each small segment 302 is reprojected into Radon space. As shown in process block 333, all pixels in the 3D FOV 300 are set to zero except for pixels in the reprojected segment 302. The segments are then reprojected by conventional methods to form the set of projected views shown in data block 304 of FIG. Preferably, it is reprojected along many different view angles that are practical from a processing time standpoint. This reprojection process continues until all individual segments 302 of the image frame are reprojected as indicated by decision block 335 to produce a corresponding projection view data block 304.

  Next, each of the reprojected segment data sets 304 is converted into a final image segment 306 using a highly constrained image reconstruction method (HYPR) as shown in process block 337 and represented in FIG. Reconfigured. This is the method described in the above-mentioned co-pending US patent application Ser. No. 11 / 482,372, which is described in more detail below with respect to FIG. It is important to note that in this case the image frame segment projection data set 304 is very sparse. This is because all signals to be reprojected are signals from only one small segment 302 of the FOV being acquired. As shown in decision block 339, when the final image frame segment 302 is backprojected, all small reconstructed final segment images 306 are combined, as shown in process block 341 and represented in FIG. A single 3D image frame 308 is formed. This is nothing but placing each final segment image 306 in the correct position within the 3D FOV 308.

  Each image acquired during the test may be reconstructed in this way. As indicated at decision block 343, the process ends when the final image frame is reconstructed.

  A possible variation of the above method is to store the final image segment 306 of each reconstructed image frame separately. As a result, the selected segments 306 can be individually displayed and focused on a small portion of the image frame FOV 308 so that diagnosis can be performed. This is particularly useful, for example, when an MIP image is generated and other vessels overlap and cover a particular vessel of interest within one image segment 306.

With particular reference to FIG. 14, the highly constrained reconstruction of each image frame segment 306 in the above-described method is described with respect to Equation (2) and illustrated in FIG. More particularly, the projection P of the image frame segment 304 is normalized, as shown in process block 231. Each projection P is normalized by dividing this projection by the projection P _c in the composite image at the same view angle. The normalized projection P / P _c is then backprojected to the FOV. This is a standard backprojection but not filtered.

  As shown in process block 233, the resulting backprojection value is added to the reconstructed image frame segment 306 and a test is made at decision block 235 to reverse all projection views of the current image frame segment 304. Determine if it has been projected. If all the projected views have not been backprojected, the next projected view in the current image frame segment 304 is backprojected, as shown in process block 237.

  Once all the projected views have been backprojected and summed for image frame segment 304, the summed image frame segment is multiplied with the composite image as shown in process block 239 to produce the final image frame segment 306. . This is a matrix multiplication, where the pixel value in the image frame segment is multiplied by the value of the corresponding pixel in the composite image. Another method for performing this highly limited image frame segment reconstruction is described in co-pending US patent application Ser. No. 11 / 482,372 (filed July 7, 2006, entitled “Highly Limited”). It is clear that it can be used as described in “Highly Constrained Image Reconstruction Method.” The contents of the above patent application are incorporated herein by reference.

  Although the present invention has been described in connection with 2D or 3D image reconstruction of k-space projection views acquired using an MRI system, the present invention can be used for radon space such as X-ray CT, PET and SPECT imaging systems. Those skilled in the art will appreciate that other medical imaging systems that acquire data are equally applicable. The present invention is not limited to k-space data acquired as a radial projection, but can be applied to other k-space sampling trajectories. For example, an image acquired using an interleaved spiral trajectory or Cartesian trajectory can be relatively changed and subjected to the same calculation as described above.

  All of the above-described embodiments of the present invention can be clearly characterized as an image reconstruction method in which a limited set of object views is acquired using a medical imaging system and images are generated from these views. The present invention is not limited to image reconstruction, but can also be applied to improving the quality of existing images. More specifically, the high SNR of the effective composite image can be incorporated into an effective low quality image frame by employing the highly limited image processing of the present invention. Image frames may be pre-acquired and reconstructed, and valid high-quality composite images are obtained by combining image frames, or by acquiring high-quality images using the same imaging system, or It may be generated by acquiring high quality images using different imaging systems of the same or different modalities.

  With particular reference to FIGS. 1 and 16, highly limited image processing is performed on a series of image frames 2. As each image frame 2 is acquired as indicated by process block 400, each image frame is stored and the composite image 3 is updated with a copy as indicated by process block 402. The composite image 3 is cumulative with a preselected number of other acquired image frames 2 and the current image frame 2. Accumulation is a matrix addition of corresponding pixels in the 2D image frame 2 divided by the number of image frames contributing to the accumulation. The result is a composite image 3 with an increased SNR that is proportional to the preselected number of image frames 2. For example, if 36 2D image frames 2 are accumulated, the SNR is 6 times the SNR of a single 2D image frame 2. The number of image frames used to generate the composite image depends on the particular clinical procedure being performed.

  As indicated generally at 404, the next step is to generate a normalized weighted image using the current 2D image frame 2 and the updated composite image 3. There are many different ways to perform this step, but the preferred method is shown in FIG. More specifically, the updated composite image 3 is “smoothed” by filtering, as indicated by process block 406. Filtering is a convolution process in which the updated 2D composite image array 3 is convolved with the filter kernel as described above. The choice of kernel size should be chosen so that when smoothing is performed, the kernel does not contain a lot of information from outside the object of interest (eg, blood vessels). If the filter kernel is significantly larger than the object of interest, the signal within the object is averaged, but its shape does not change. On the other hand, if the filter kernel is smaller than the object of interest, its shape, or profile, is smoothed.

Still referring to FIG. 16, the current 2D image frame 2 is also smoothed or filtered as indicated at process block 408. That, 2D image frame array 2 is convolved with the filter kernel to perform a number of functions lowpass filter-ring. Normalization is then performed by dividing the pixel value in the filtered current image frame (T) by the corresponding pixel value in the filtered composite image (C _t ), as shown in process block 410. A weighted image (T _w ) is generated.

As indicated at process block 412, a highly limited (HYPR) image frame 4 is then generated. This image frame 4 is generated by multiplying the updated composite image array 3 and the normalized weighted image (T _w ). This is a multiplication of the corresponding pixel values in the two images. Next, as shown in process block 414, the resulting 2D HYPR image 4 is output to the display and the system loops back to obtain and process the next 2D image frame 2. When this procedure is complete, the program ends as determined at decision block 416.

As mentioned above, there are many alternative ways to generate a normalized weighted image (T _w ). Two of these methods are illustrated in FIGS. With particular reference to FIG. 17, the first alternative method includes a first step, indicated by process block 418, in which the acquired 2D image frame array 2 is divided by the updated composite image 3. This is to divide each pixel value in the acquired 2D image frame array 2 by the corresponding pixel value in the updated composite image array 3. The resulting 2D image frame resulting from the division is then smoothed or filtered as shown in process block 420 to produce a normalized weighted image (T _w ). This filtering operation is the same convolution process described above in connection with process blocks 406 and 408.

Another alternative method for generating a normalized weighted image (T _w ) is shown in FIG. The method converts the acquired 2D image frame 2 to Radon space by capturing projected views of the image from different view angles, as shown in process block 422. As shown in process block 424, the updated composite image 3 is also converted to Radon space by calculating a projected view in the same set of view angles used to convert the 2D image frame 2. The As indicated at process block 426, the image frame projection view P is then normalized by dividing by the composite image projection view P _c . This is a division of the corresponding elements in projections P and P _{c at} the same view angle. Next, at process block 428, a normalized weighted image (T _w ) is generated by backprojecting the normalized projection (P / P _c ) in a conventional manner. This is a direct backprojection, not a filtered backprojection.

It is the figure which showed application of this invention to a medical imaging use. FIG. 6 is a diagram of how k-space is sampled in a typical Fourier or spin warp image acquisition using an MRI system. FIG. 6 is a diagram of how k-space is sampled in a typical projection reconstruction image acquisition using an MRI system. FIG. 2 is a graphical representation of a conventional backprojection step in an image reconstruction process. Figure 3 is a graphical representation of a highly limited 2D backprojection step according to the present invention. FIG. 3 is a graphical representation of highly limited 3D backprojection according to the present invention. 1 is a block diagram of a magnetic resonance imaging (MRI) system used to implement the present invention. 7 is a pulse sequence used in the MRI system of FIG. 6 to implement an embodiment of the present invention. FIG. 8 is a graphical representation of k-space data sampled using the pulse sequence of FIG. 8 is a flowchart of a preferred embodiment of the present invention used in the MRI system of FIG. 6 using the pulse sequence of FIG. FIG. 10 is a flowchart of one preferred method for reconstructing a HYPR image frame that forms part of the method of FIG. FIG. 11 is a flowchart of another preferred method for reconstructing a HYPR image frame. FIG. 12 is a pulse sequence used in the MRI system of FIG. 6 to implement another embodiment of the present invention. FIG. 13 is a flowchart of a preferred embodiment of the present invention used in the MRI system of FIG. FIG. 14 is a flowchart of the highly limited image reconstruction method used in the method of FIG. FIG. 15 is a diagrammatic representation of the process shown in FIG. FIG. 16 is a flowchart of another preferred embodiment of the present invention. FIG. 17 is a flowchart of an alternative method used in the procedure of FIG. 18 is a flowchart of another alternative method to the method used in the procedure of FIG. FIG. 19 shows a reconstructed image of the entire FOV.

Claims (4)

A highly limited reconstruction method for acquired image frame datasets,
And reconstructing the acquired image frame data set or al picture image frame,
Dividing the front Kiga image frame into a plurality of segments,
Reprojecting each segment of the plurality of segments to generate a corresponding plurality of segment projection view data sets;
The composite image containing a priori information about the object to be imaged is reconstructed from the acquired image frame data set, using said segment projection view data sets corresponding in backprojection process that is highly constrained Reconstructing a final image frame segment, and a highly limited reconstruction method.   The method of claim 1, wherein the final image frame segments are combined to produce a final image frame.   The method of claim 1, wherein the final image segment is selectively displayed.   The method of claim 1, wherein the composite image is reconstructed using the acquired image frame data set combined with other acquired image data.

Images